Source
Rasooly and Rozzi (2025), arXiv:2503.03312, LIT-958, read in full in NOTE-778.
What was actually shown
- The experiment (§§4–5). 817 binary markets on Manifold were each
assigned at random to a yes trade (+5 points), a no trade (−5 points)
or control, and their hourly prices were recorded for 30 days with a
snapshot at 60. The yes–no gap β̂1 was:
- 0.0748 at 1 week, about 25% reversion (Table 3);
- 0.0686 at 30 days, with 25% reversion in the first week and 6% in the next three;
- 0.0586 [0.0114] at 60 days in the 771 unresolved markets, about 41% reversion (Table 2). All have p < 0.01. Effects were symmetric between yes and no. They survived dropping possibly related markets (Appendix C) and appeared, with faster reversion, in 110 real-money-convertible markets (§6).
- The mechanism in the model (§2, Appendix A). Traders are finitely many, each with finite wealth and risk aversion (DARA), and trade against a constant-product market maker. Lemma 3 proves every trader responds to an upward shock by pushing the price down. Simulations show the response is partial: with 11 traders and no learning, about 60% of the shock remains once prices settle, and reversion slows as the price approaches its old value. More traders (82% reversion at m = 60) or a common prior (52%) undo more. Learning from the price undoes less, down to none at λ = 1 (Table A1).
What could have come out otherwise: the received view (Wolfers and Zitzewitz) predicts reversion within a short transition, as Camerer found at the racetrack within about 20 minutes. The randomised design gives the counterfactual path that historical studies lacked.
What this does not say
- Not that traders came to believe the shocked price. The model gives persistence even when no belief moves (λ = 0). The data cannot distinguish that from learning. Whether the price that persists is anyone’s belief is open.
- Not that a price is a group belief. A price aggregates positions weighted by wealth and risk attitude. The claim is about an aggregate estimate departing from what members’ information supports, not about a belief the community holds.
- Not that trader count, volume or outside information cause lower manipulability. Those differences are median splits on traits that were not randomised, at p between 0.04 and 0.19, and the authors flag this.
- Not shown for large real-money markets. Manifold runs mainly on play money redeemable as charity, and its user base is unusual. The real-money check is small (110 markets, about 6 days, no control). The authors predict more persistence in high-value markets, untested.
- Not about other shock sizes. The authors expect very large shocks to be less persistent, since they are not “credible”. One size was tested.
- Not “up to 60 days”. Sixty days was the last measurement, not where the effect ended.
Connections
- QUESTION-047. A measured case of a community-level estimate holding
a value that no member reasoned to, because the value was set at
random. It bears on the question’s first clause in a weakened form:
“settle on” an estimate rather than a belief. The question’s second
clause is answered in the model (number of traders, prior agreement,
learning rate), but only correlationally in the field. No
answers:relation is declared, because whether any member holds the persisting value is exactly what is unknown. - THEORY-242 (spiral of silence). It shares the shape: an aggregate signal departs from what members hold, and members then read it as information. There the signal is expressed opinion; here it is price. In both cases the record lacks an empirical measure of members’ private states.
Note of 2026-10-20: narrowed to the measurement
Promotion audit, batch 050, under ADR-047 v3. The study is one pre-registered randomised field experiment, read in full with its regression tables (Tables 2, 3, A1, A3 and A5).
Kept in the title:
- The persistence measured on Manifold. β̂1 was 0.0748 at one week, 0.0686 at 30 days, and 0.0586 [0.0114] at 60 days in the 771 markets not yet resolved, all at p < 0.01. So about 59% of the shock remained.
- The slowing of reversion.
- The model’s mechanism, reported as the model’s.
Dropped from the title:
- The generality of “a prediction-market price”. This is one platform, at play money, with one shock size, and the small real-money follow-up reverted faster.
- The causal “because”. Whether the persisting price is anyone’s belief is the open part, together with replication elsewhere, in QUESTION-212.
Note of 2026-10-20: the wider claim, filed Proposed
Under ADR-047 v4 the narrowing-loss sweep (batch 071) filed THEORY-810 as Proposed, extending this entry. It carries the general reading this entry’s title carried before the promotion audit narrowed it: that prediction-market prices in general keep much of an unsupported shock for weeks, as Rasooly and Rozzi propose. This entry is its evidence and is not changed by the filing.